Decision and game theory
Chapter1: introduction
4 board sources of difficulty: complexity, uncertainty, multiple objectives, competing
viewpoints.
● complexity: too many uncertain scenarios, too many factors, too many alternatives
(sequential), too many consequences.
● uncertainty: future events beyond the DM control may result in good or bad
consequences; deman, consumer attitude, market conditions, competition, evolution
of price.
● Multiple objectives: quality vs cost, short term vs long term profit,economic vs
environmental benefit, stakeholders’ vs employees’ benefit, expected profit vs risk.
⇒ decision analysis provides both a framework and specific tools for dealing with such situations.
● competing viewpoints: from a single perspective, changes in the decision criteria may
change the optimal decision policy
many decision makers⇒ different perspectives( priorities ranking may differ)
cognitive illusions: humans are often very certain of knowledge that is false (rational).
structuring decisions steps:
1. the objectives: specific goal to reach (measurable to evaluate the alternatives,
dictates the way to measure the results and the type of uncertainty) ( ultimate goal/
objective)
2. the decision alternatives: accept/ reject (clear), some of them need some
analysis( wait and see, do nothing, pay for safety or insurance)
3. the state of nature: identified through the level of knowledge of the state of
nature:
a) uncertainty: extreme case; DM controls without knowledge of the likelihood
of their occurence.
b) risk: possible to use available information to assess the probability
distribution of the occurrences of situations out-of the control of DM.
c) certainty: extreme case: all outcomes of the decision problem are known
beforehand.
4. the outcome (gain or loss) function
5. the environment of the decision problem
Solving the decision problem:
By identifying the optimal set of actions to take is usually obtained through backward
iterations; calculations (easy through softwares).
sensitivity analysis: to answer “what if” question.
Any minor change in the problem affects the solution⇒ re-evaluate sensitive parameters.
⇒ help the DM to redefine the objectives or to consider alternatives leading to more structured model
⇒ more adequate decision model⇒ more pertinent solution.
simple and sequential decision problems:
1. simple: one decision⇒ one alternative (chosen form a set of alternatives)==> direct
consequence.
2. Sequential: require a series of interrelated decisions ⇒ DM select an alternative from each
subset of decision alternatives⇒ resulting gain depends on the entire sequence of actions
taken.
payoff tables:
Payoff: the consequences resulting from a specific combination of a decision alternative and
a state of nature.( profit, cost, time….)
Payoff table: shows all payoffs for combinations of decision alternatives and states of
nature.
Decision tree:
is a graphical representation that structures in a chronological order the dcision alternatives
and the states of nature.
● nodes: represented by squares
● probabilities; represented by circles
Chapter 2: Decision under uncertainty:
4 types of decision making attitudes under certainty:
1. Optimistic DM: max-max or min-min
2. Pessimistic DM ( Wald’s criterion): max-min or min-max
3. Moderately optimistic DM (Hurwicz criterion):alpha*max+(1-alpha)*min
4. Rational DM ( Laplace criterion): mean: average payoff over all entries
Chapter 3: Decision under risk:
Decision under risk without additional information
3 possible criteria can be used:
1. Maximum likelihood Criterion
2. Expected Monetary Value (EMV)
3. Expected-Variance Criterion
A. Maximum likelihood Criterion:
the states of nature have probabilities of occurrence; the DM identifies the most likely
event, ignores others and picks act with the best payoff.
B. Expected Monetary Value (EMV):
the DM calculates for each alternative the weighted average payoff, then selects the
act providing the best expected value.
C. ExpectedValue-Variance Criterion:
the DM optimizes both the expected value and the variance. The DM maximaizes
(E(Z)-K*variance(Z)).
.
Decision under risk with additional information:
Bayesian analysis/ a posteriori analysis: Look for additional information about the states of
nature in order to revise the initial distribution.
imperfect information: does not give the exact state of nature that would happen ⇒ it is given in terms
of probability distribution.( prior probabilities⇒ additional info ⇒ A posteriori probability).
May arise from a market survey, a consulting advice, a study, a forecast, research work,
expert....
⇒ To get the right answer we need to calculate the EVII.
Bayes’ theorem:
Additional information reduces risk in decision making. It can in some extreme situations
completely remove the risk and provide a certain environment.
Evaluate the value of perfect information (EVPI):
If the state of nature that will occur is known with certainty, then the best alternative can be
determined with certainty as well⇒ could give the best value for decision, which can be then
compared to the value expected under current information.
The difference between them is THE VALUE OF ADDITIONAL INFORMATION.
EVPI= expected value(payoff) under certainty- expected value(payoff) under
uncertainty.
EVPI places an upper bound on what one would pay for additional information
Chapter 4: Utility functions
expected value criterion advantage: simplicity by summarizing all the available information on
one score⇒ His main weakness; the criterion loses much of it which can make its application
dangerous.
Several real situations require the consideration of intangible factors that are difficult to
express by a simple measure of the EV.
Von Neumann-Morgestern axioms:
Notations and terminology:
A lottery L, L’ and L” are any uncertain event having random outcomes.
Transitivity Axiom:
L>L’, L’>L” then L>L”
Independence Axiom:
For all lotteries p€[0,1], we have L>L’ ⇒ pL+(1-p)L”> pl’+(1-p)L”
Continuity Axiom:
if L>L’>L” ⇒ there is p€[0,1] such that L’=pL+(1-p)L”
⇒ An agent accepting all four axioms above is called Von Neumann-Morgestern rational or just
VNM-rational.
VNM Theorem:
L>L’ if only U(L)>U(L’)
U(x*)=1⇒ best outcome and U(x0)=0 ⇒ the worst outcome.
X* with probability p, and X0 with probability 1-P
U(x)=p.x ⇒ certainty equivalent (CE) of the lottery L.
Curve: (x;U(x)).
DM attitudes toward risk:
Risk averse: E(L)> CE(L) ⇒concave⇒ RP(L)=E(L)-CE(L)> 0 ⇒ the risk premium.
⇒ RP(L) is the amount of money that the DM is willing to pay to avoid risk.
Risk neutral: E(L)=CE(L) ⇒ RP= 0.
⇒ linear.
Risk taker: E(L)<CE(L) ⇒convex⇒ RP(L)=E(L)-CE(L)< 0
⇒ the risk premium is the amount of money the DM is willing to receive in addition to the expected
value of the lottery to give up risk.
⇒ the higher the curvature the more averse or seeker the DM.
⇒ The same DM may have different risk attitudes based on the problem situation or the values
considered.
Chapter5: Game Theory
Game theory studies the behaviors and decisions of agents placed in a situation of
interdependence.
main originality: players are aware of their goals and other protagonists goals as well.
DM should be rational and pursue exogenous objectives.
Strategic situation: where you should think ,about what depends on what others do.
Game theory helps in examining and predicting how people behave in a strategic situation.
we have 3 main elements in game theory:
Players:
Strategies: simple or complex ( well defined).
Payoffs: final returns to the players that are measured in terms of utility obtained by the
player.
Goal: player wants to maximize his utility.
Players are rational:
assumptions: Humans are rational beings
Humans seek the best alternative in a set of possible choices.
WHY?: Narrow down the range of possibilities.
Predictability.
Types of Games:
cooperative: players can make irrevocable agreements with each other.
Non-cooperative: impossibility of communication or a prohibition of consultation between
competitors.
Or, it is assumed that players are in no way compelled to take any action respecting the
agreements that they may have previously concluded between them.
● sequential vs simultaneous moves
● Single play vs lterated
● Zero vs non-zero sum
● perfect vs imperfect information
Non-Cooperative Games:
The children’s hand game:
each player simultaneously makes a figure of rock, paper or scissors
Outcome: rock dominates scissors, scissors dominates paper, and paper dominates rock.
- In a 2 person game, the player who makes the dominating figure wins the game.
- when both make the same figure, it’s a draw and neither players win.
who reveals first has a disadvantage, since the second player will choose an action that
results in a win.
when both players reveal their actions simultaneously, neither of them has any prior
information on the actions of the other player.
Battle of the Sexes:
wife: wants to go shopping
husband: wants to watch football.
Both want to spend time together
Prisoners’ Dilemma:
Police arrested 2 suspects for a crime:
want them to confess.
Deal offered: One confess and the other
remain silent; 1 wil be set free and the other
will be punished with 15 years.
Both defect to confess; punishment 10 years of
prison for both.
Both remain silent; One year sentence.
Dominance:
Notations:
Definitions:
Equilibrium:
If one player has a strictly dominant
strategy: he chooses it without any
hesitation since his best choice is
independent of that of the other players.
If all players have a strictly dominant strategy: An Equilibrium of dominant
strategies exists that is determined without having to consider the behavior of the
other players.
The existence of an equilibrium in weakly dominant strategies is almost miraculous.
In general, a player must consider other players’ strategies. He may then reduce his
set of strategies based on rational behavior.
By assuming that the dominated strategies will never be played, we can eliminate
them.
If this elimination process converges towards a single outcome, everyone can
unambiguously anticipate the behavior of the others.
⇒ The obtained outcome is called equilibrium by elimination of strictly dominated strategies.
⇒ Rarely exists. The elimination process often stops while the players still have several
alternatives, none of them being strictly dominated.
⇒ The process of elimination can be multiple.
The process can converge to several possible outcomes according to the order in
which the eliminations succeed, but there is no reason to assume that players follow
the same order in the elimination process.
Most games are not solvable by iterated strict dominance.
⇒ Nash Equilibrium: a set of strategies, one for each player, that are each best responses
against one another.
A Nash equilibrium is where the strategy of one player is the best strategy when
other players play their best strategy.
⇒ A Nash equilibrium is a situation where no player has an interest in deviating (alone) from
the situation obtained.
⇒ Not all games have a pure Nash equilibrium.
⇒ Every game has a Nash equilibrium, and some games may have several.
Simultaneous Move Game:
“Static” game: players take their actions at the same time, and the game ends immediately. ⇒
Game theory that the outcome is a Nash equilibrium.
The battle of the sexes: has a Nash equilibrium.
The children’s hand game: has no pure Nash equilibrium.
The prisoner’s Dilemma: has a Nash equilibrium.
Two players zero-sum games:
Zero-sum games: for which player strategies can not increase or decrease available
resources. The sum of the gains is constant and therefore if the gain increases for
one of the players it decreases inevitably for the other player.
⇒ The battles of the sexes game or the prisoner’s dilemma are non-zero sum games.
A game is pure strategy, if the optimal strategy for a player is the same whatever the
game of the 2nd.
In pure strategy, the first player selects the maximum among the identified values. In
other words the 1st player maximizes the minimum allowed gains(offensive).
However, the 2nd player minimizes the maximum loss (defensive).
⇒ when there is a stable solution, both players should stick to their min-max and max-min
solutions respectively.
⇒ Saddle point doesn’t always exist.
⇒ when the game is repeated, a player needs to change strategy.
when the game
doesn’t have a saddle point, game theory advises each player to assign a probability
distribution over his set of strategies.
Pi= probability that player 1 will choose strategy i (i=1,2,...m).
Qj= probability that player 2 will choose strategy j (j=1,2,...n).
